Can I read more about this somewhere? On Haswell I've observed a 5% speedup (for a whole program that mostly does a bunch of other work) from spacing a bunch of arrays that are typically all pretty small and all accessed consecutively all the way through to have their start positions differ by some amount of cache lines other than a power of 2. So it certainly seemed to me like I was working with a set-associative cache where power of 2 strides were pathological.